Classification of breast mass lesions using model-based
analysis of the characteristic kinetic curve derived from
fuzzy c-means clustering
5
Yeun-Chung Chang1, Yan-Hao Huang2, Chiun-Sheng Huang3, Pei-Kang Chang2, Jeon-Hor
Chen4,5, and Ruey-Feng Chang2,6
1
Department of Medical Imaging, National Taiwan University Hospital and National Taiwan
University College of Medicine, Taipei, Taiwan
10
2
Department of Computer Science and Information Engineering, National Taiwan University,
Taipei, Taiwan
3
Department of Surgery, National Taiwan University Hospital and National Taiwan University
College of Medicine, Taipei, Taiwan
4
Department of Radiology, China Medical University Hospital, Taichung, Taiwan
15
5
Tu and Yuen Center for Functional Onco-Imaging and Department of Radiological Science,
University of California Irvine, California, USA
6
Graduate Institute of Biomedical Electronics and Bioinformatics, National Taiwan University,
Taipei, Taiwan
Manuscript type: Original Research
20
Running Title: Fuzzy C-means Clustering in DCE-MRI
*Correspondence Address:
Professor Ruey-Feng Chang, PhD,
Department of Computer Science and Information Engineering
National Taiwan University,
25
Taipei, Taiwan 10617, R.O.C.
Telephone: 886-2-33664888~331
Fax: 886-2-23628167
E-mail: rfchang@csie.ntu.edu.tw
30
Professor Jeon-Hor Chen, M.D.,
Center for Functional Onco-Imaging, University of California Irvine,
No. 164, Irvine Hall, Irvine, CA 92697, USA
Tel: 1-949-824-9327
Fax: 1-949-824-3481
35
E-mail: jeonhc@uci.edu
Acknowledgment
The authors would like to thank the National Science Council of the Republic of China for
40
financially supporting this research under Contract No. NSC 97-2221-E-002-166-MY3.
Classification of breast mass lesions using model-based
analysis of the characteristic kinetic curve derived from
45
fuzzy c-means clustering
Abstract
The purpose of this study is to evaluate the diagnostic efficacy of the representative
characteristic kinetic curve of dynamic contrast-enhanced (DCE) magnetic resonance imaging
50
(MRI), extracted by fuzzy c-means (FCM) clustering for the discrimination of benign and
malignant breast tumors using a novel computer-aided diagnosis (CAD) system. About the
research dataset, DCE-MRI of 132 solid breast masses with definite histopathologic diagnosis
(63 benign and 69 malignant) were used in this study. At first, the tumor region was
automatically segmented using the region growing method based on the integrated color map
55
formed by the combination of kinetic and area under curve (AUC) color map. Then, the fuzzy
C-means (FCM) clustering was used to identify the time-signal curve with the larger initial
enhancement inside the segmented region as the representative kinetic curve and then the
parameters of the Tofts pharmacokinetic model for the representative kinetic curve were
compared with conventional curve analysis (maximal enhancement, time to peak, uptake rate
60
and washout rate) for each mass. The results were analyzed with a receiver operating
characteristic (ROC) curve and student’s t-test to evaluate the classification performance.
Accuracy, sensitivity, specificity, positive predictive value, and negative predictive value of
the combined model-based parameters of the extracted kinetic curve from FCM clustering
were 86.36% (114/132), 85.51% (59/69), 87.30% (55/63), 88.06% (59/67), and 84.62%
65
(55/65), better than those from a conventional curve analysis. The AZ value was 0.9154 for
Tofts model-based parametric features; better than that for conventional curve analysis
(0.8673) for discriminating malignant and benign lesions. In conclusion, model-based
analysis of the characteristic kinetic curve of breast mass derived from FCM clustering
provides effective lesion classification. This approach has potential in the development of a
70
CAD system for DCE breast MRI.
Key Words: DCE-MRI; breast; pharmacokinetic; color map; AUC; kinetic;
1. INTRODUCTION
Magnetic Resonance (MR) of the breast is the most sensitive tool to detect breast cancer [1-2].
75
Interpretation of breast MR requires not only a focus on morphologic changes but also on the
pattern of the areas with increased enhancement [3-5]. In view of the tremendous amount of
three-dimensional (3-D) imaging data provided by a current state-of-art MR scanner, the
requirement for computer assistance is increasing in order to avoid human error by the
interpreting radiologist. The time-signal intensity curve (TIC) from dynamic contrast
80
enhanced (DCE) MR imaging has been used as an effective tool to determine the possibility
of malignancy in addition to the morphologic features [1,3-5]. A rapid upslope and quick
wash out pattern in TIC on DCE-MRI has been widely accepted as an important parameter for
predicting the possibility of malignancy. However, the selection of a region of interest (ROI)
as the representative area of the tumor is operator-dependent. Some automatic computer
85
assistance programs have the capability of classifying the TIC of each voxel for a whole 3-D
DCE breast MRI study and to alert the interpreting radiologist to possible malignancies, thus
helping to avoid unnecessary human error [6]. These methods can effectively increase the
awareness of radiologists but lack more specific characterization of particular lesions. There
are, however, some problems with this approach, including false negative due to the adoption
90
of a minimum threshold of enhancement [6] and inaccuracy of TIC pattern due to motion [7].
For quantitative analysis of a tumor in DCE-MRI, the majority of prior studies focused
on four conventional curve parameters of the TIC [7-8] (maximum enhancement, time to peak,
uptake rate, and washout rate), rather than using a pharmacokinetic model to fit the TIC
[9-12]. It has been shown that the initial area under the gadolinium curve (IAUGC) is a mixed
95
parameter that can display correlation with pharmacokinetic parameters [13]. Kinetic color
mapping [14] can highlight area with greater enhancement in early phase and thus
increase detection rate of occult breast cancers. IAUGC is considered associated with
physiologic meaning with lack of assumption and ease of implementation [13]. It is
hypothesized that a combination of kinetic color mapping and area-under-the-curve
100
(AUC) can be potentially useful to find enhancing area with greater clinical
significance. The fuzzy C-means (FCM) clustering algorithm is an unsupervised
-1-
clustering technique and useful for image segmentation and pattern recognition [15].
The representative kinetic curve by FCM has been successfully applied in DCE breast
MRI and shown better than using curve by averaging the entire lesion [7-8]. In
105
addition, pharmacokinetic model of DCE MRI is not only being used increasingly to
noninvasively monitor the action of antiangiogenic and antivascular therapy [16] but
also helpful in differentiating benign from malignant breast cancers [17]. In this study,
we used the TIC acquired from DCE-MRI with a kinetic color map [14] and
area-under-the-curve (AUC) analysis [13,18] followed by a region growing method [16] for
110
tumor segmentation, and then using the fuzzy C-means (FCM) clustering technique [17] to
produce a representative TIC of the targeted lesion. Each representative TIC derived from
FCM clustering was then fitted by using a pharmacokinetic model and compared with the
results of a conventional curve analysis. The purpose of our study was to evaluate the
accuracy of tumor classification with the information from the TIC with this novel computed
115
aided diagnosis (CAD) system.
2. MATERIALS AND METHODS
2.1 Patients
120
In this study, we used the MR dataset of 99 consecutive patients between August 2006 and
September 2009. A total of 132 mass lesions (63 benign and 69 malignant, size range from 0.7
to 8.5 cm, 2.33±1.84cm), diagnosed by three breast radiologists (3, 3 and 8 years of
experience in interpreting breast MRI) using BI-RADS lexicon in 3-D DCE-MRI, in 82
patients (age range, 32 to 85 years; mean ± standard deviation, 53.24±9.82 years) were used
125
to evaluate the performance of our computer aided diagnosis (CAD) system. All the 132
breast lesions had clinical impression of breast mass based on image findings of
mammograms or ultrasound and all had the final histological proof through pathological
examination of tumor tissue specimen obtained from core needle biopsy or surgical resection.
None of them received breast MRI for screening. The pathological diagnosis was made by
-2-
130
each in-charge pathologist who had at least 3 years of clinical experience. The final diagnosis
of these breast tumors included invasive ductal carcinoma (n=51), invasive lobular carcinoma
(n=3), ductal carcinoma in situ (n=15), fibroadenoma (n=19), papillomas (n=6) and focal
fibrocystic change (n=38). This study was approved by the Institutional Review Board, and
informed consent was waived for our retrospective study.
135
2.2 DCE-MRI Imaging
All DCE-MRI studies were acquired with a 1.5T MR scanner (Signa Excite HD, GE
Healthcare, Milwaukee, WI, USA) with dedicated 8-channel breast coils in the prone position.
The dynamic study with bilateral whole breast coverage was performed with the following
140
parameters: fat suppressed 3D fast spoiled gradient echo (FSGR), TR/TE/TI = 3.5/1.7/14 ms,
flip angle 12 degrees, matrix 256×160, image size 256 × 256 pixels, slice thickness 2-2.5 mm
without gap, acquisition 0.75, and field of view 24×24 to 30×30 cm. There were a total of 35
acquisitions for the DCE-MRI study. Each acquisition included 56 axial slices and covered
11.2-14 cm distance in cranial-caudal (Z-axis) direction. The temporal resolution of
145
DCE-MRI was 18-20 seconds. Intravenous injection of MR contrast agent (CA) (0.5 mmol/ml,
Gadodiamide, Omniscan, GE Healthcare; Magnevist, Bayer-Shering Pharmaceuticals) was
performed with a bolus injection (flow rate 4 ml per second) simultaneous with the beginning
of the acquisition and followed by saline flushing.
150
2.3 Conventional Time-signal Curve Analysis
Conventional kinetic analysis of the TIC in all studies was performed by three in-charge
breast radiologists with the information of clinical history and other breast imaging findings
using the software (FuncTool 3.1.01, GE Healthcare, Milwaulkee, USA) in a commercially
available workstation. The ROI was placed at the area with most intense enhancement in the
155
suspicious lesion [3]. Usually, multiple ROIs of a lesion were obtained and the most
characteristic or suspicious ROI was used to make a conclusion. At least 3-5 pixels were used
-3-
for small enhancing lesions. For large lesion, the most enhancing part of the tumor was
selected.
160
3. FCM Clustering of Pharmacokinetic Model
There were two major steps, 1) tumor extraction and, 2) curve identification, for
obtaining characteristic curve of each targeted mass lesion identified on DCE MRI. The
whole procession time, including manual selection of the interested tumor area, was about 90
seconds.
165
3.1 Tumor Extraction
The first step consisted of a tumor extraction algorithm performed by finding the
intersection of a kinetic color map [14] and an AUC color map [15] for the whole breast
(Figure 1). The kinetic color map was obtained from categorization of TIC according to
170
relative enhancement (RE) ratio. The AUC color map was generated from the relative
accumulation of contrast enhancement on TIC. The concept was to obtain the most
enhancement region representing functioning part of the tumor on the integrated color map.
This approach could improve the performance of our system. Therefore, a specific intensely
enhanced area with well defined margin could be extracted. After reviewing the DCE MR
175
images and the integrated functional map, only one seed was manually placed in the target
mass lesion within a volume of interest (VOI) which included the whole tumor region in the
3-D spatial domain on the integrated functional color map. Because some enhanced normal
tissues would be connected to the target mass lesion, the proper VOI could assist in excluding
these tissues for correct segmentation. Finally, a 3-D tumor segmentation was obtained using
180
a region growing method [16] (Figure 1 and Figure 2).
Because malignant lesions tend to have a RE ratio greater than 100% in the early phase
[3], we used 50%, 100% and 200% enhancement as cutoff points for kinetic color map. After
evaluating the contrast-to-noise ratio of segmented targeted mass lesions, we assigned three
-4-
colors for representing different ranges of relative kinetic enhancement: 1) yellow for a RE ≥
185
200%, 2) red for a RE ratio < 200% but ≥ 100%, 3) blue for a RE ratio < 100% but ≥ 50%.
The AUC color map was used to find the largest cumulated signal intensity over time on
the TIC of each voxel of the segmented mass lesion. Different colors (red, yellow and blue)
were used to display the larger values of the AUC color map based on cumulative histogram.
The thresholds, 90%, 80% and 60%, were chosen after reviewing all processed data in which
190
all mass lesions in our study group showed AUC value larger than 80%. Because most tumors
had larger AUC value (≥90% of whole distribution in the cumulative histogram), they
maintained a greater cumulative enhancement which was obviously different from neighbor
tissues. However, some lesions with smaller AUC value (≥80% and < 90% in the cumulative
histogram) were difficult to separate from surrounding normal tissue. Hence, the regions with
195
AUC value larger than 80% were more appropriate to use as the threshold for segmentation.
Moreover, the some normal tissues were enhanced with middle AUC value between 80% and
60%, it was assigned as blue region for visualization and confirmation of the clinic
examination. No color was assigned if the AUC was less than 60%. Therefore, mass like
lesions were marked by yellow and red. For the integrated color map, purple region is both
200
red in the AUC color map and yellow in the kinetic color map. Besides the purple region, the
red region in the integrated color map is red in the AUC color map, the yellow region is
yellow in the kinetic color map.
3.2 Curve Identification and Analysis
To obtain the specific and characteristic information from the targeted tumor from the
205
segmented VOI, the FCM clustering technique [17] was applied to find the most characteristic
and significant curve that fitted the TICs of all pixels in the segmented tumor. In the previous
study [7], manual definition of the VOI containing the tumor region was applied first.
Tumor region was segmented by the FCM clustering technique and then the maximum
enhanced curve was picked up by the FCM to extract four conventional features for analysis.
210
In contrast, integrated color map of the whole breast was built first and the targeted regions
-5-
were highlighted by the characteristic of tissue enhancement for segmentation in our study.
Moreover, the only one representative TIC (cselect) was extracted from FCM selection
function to represent the characteristic of the selected mass, the selection function was defined
as
215
ck1  ck 0
k 1,2,...,C
ck 0
cselect  arg max
(1)
where ckt is the intensity of center ck at tth time point, and C is the number of clusters. Then
the representative TIC was then fitted with a Tofts pharmacokinetic model using
compartmental model [11-12].
The representative TIC was also analyzed using conventional curve analysis, i.e.,
220
maximum enhancement (Fk1), time to peak (Fk2) (min), uptake rate (Fk3) (min-1), and washout
rate (Fk4) (min-1) for comparison (Table 1).
For extracting the diagnosis features, the representative curve derived from FCM
clustering for each tumor was fitted with the Tofts pharmacokinetic model [11-12].
The Tofts model is defined by [11-12]
225
Ct (t )  vPCP (t )  K trans[CP (t )  e
 k ep t
]
(1)
where Ct(t) is the contrast agent concentration in the tissue with time t, vp is the fractional
volume of blood plasma, Cp(t) is the contrast agent concentration in the blood plasma with
time t, Ktrans is the volume transfer constant between the blood plasma and extracellular
extravascular space (EES), kep is the rate constant between blood plasma and EES, and  is
230
the convolution operator. The fractional volume of EES (ve) is defined by
ve 
K trans
.
k ep
(2)
Because the Tofts model requires the arterial input function (AIF) for Cp(t), in this paper
the AIF was estimated from the concentration of contrast agent concentration in the ascending
-6-
aorta. The Levenberg-Marquardt algorithm [18] which can approximate the curve to find the
235
numerical solution of nonlinear function was iteratively used to fit the nonlinear equation, and
the parameters. This approach could smooth the characteristic TIC extracted from FCM
clustering as well as reduce motion artifact. The analysis of conventional curve and
pharmacokinetic model was shown in Table 1.
A general binary logistic regression [19] was applied to classify these solid breast
240
masses based on the parameters of the pharmacokinetic model. The leave-one-out
cross-validation method [20] was used to estimate the performance of the binary logistic
regression.
3.3 Statistical Analysis
245
An unpaired Student’s t-test was used to analyze the parameters associated with benign
or malignant lesions. A p value of less than 0.05 was considered significant. The parameters
from the conventional curve analysis and pharmacokinetic models for the FCM clustering
TIC for discriminating benign from malignant were individually tested by the one-sample
Kolmogorov-Smirnov test. The overall performance was evaluated by using a receiver
250
operator characteristic (ROC) curve analysis program (LABROC1, 1993; Charles E. Metz
MD, University of Chicago, Chicago, Ill). Accuracy, sensitivity, specificity, positive
predictive value (PPV), negative predictive value (NPV), and AZ index of the ROC were used
to evaluate the diagnostic performance. The results were also compared with results of breast
radiologists’ diagnosis solely based on malignant and benign kinetic types proposed by Kuhl
255
et al. [3].
4. RESULTS
4.1 Tumor Extraction and Features
Using the intersection of the kinetic color mapping and the AUC mapping, mass-like
breast lesions with enhanced components were segmented (Figure 2). FCM clustering was
260
capable of extracting the most characteristic and representative TIC, as shown in Figure 3.
-7-
The parametric values of the conventional curve analysis and pharmacokinetic models
for the FCM clustering extracted TIC in both benign and malignant masses are shown in
Table 1. The mean value, standard deviation (SD), median value, and p-value of Student’s t
test or Mann-Whitney U test for various features were calculated. The Kolmogorov-Smirnov
265
test was applied to test for a normal distribution. If the distribution of a feature was normal,
the mean value and standard deviation were listed and the Student’s t test was used (Fk1 and
vp). Otherwise, the median value was listed and the Mann-Whitney U test was used (Fk2, Fk3,
Fk4, Ktrans, kep and ve). Before the calculation of the Student’s t-test, the Levene’s test had
been used for verifying the equality of variances. There were significant differences between
270
benign and malignant lesions using each conventional curve characteristics in Table 1.
Benign lesions were lower in maximum enhancement (Fk1) (1.233±0.681 vs. 1.589±0.452)
(p<0.001), showed a longer time to reach peak enhancement (Fk2) (465.372 vs. 180.144
seconds) (p<0.001), slower uptake (Fk3) (0.003 vs. 0.009) (p<0.001) and slower washout
(4.914×10-4 vs. 8.969×10-4) (p=0.005) compared with those of malignant lesions. The findings
275
were comparable with the characteristics of benign or malignant lesions in TIC classification
[3]. Using a Tofts pharmacokinetic model, it was also found that benign lesions had
significant lower kep (0.142 vs. 0.411) and vp values (0.415±0.162 vs. 0.600±0.215) (p< 0.001)
than malignant lesions. The Ktrans and ve value of benign lesions was significantly higher than
malignant lesions (0.860 vs. 0.652 and 5.639 vs. 1.438) (p< 0.001).
280
The groups of combined features with different curve identifications were tested by
using a binary logistic regression (Table 2). There was obviously higher accuracy, sensitivity,
specificity, PPV, and NPV if a combination of all pharmacokinetic parameters was used
(p=0.7258, 0.3690 and 0.5078, respectively) while higher specificity and PPV if using
conventional kinetic characteristics (p=0.5708 and 0.7032, respectively) (Table 2). However,
285
none of the comparison between the two methods reached statistical significance.
Each conventional curve parameters and pharmacokinetic parameter was significantly
different (Table 1), there was no significant improvement in discriminating benign from
malignant lesions when all parameters were considered together (Table 2). The scatter plots
show that there was better performance for Fk2 (time to peak) than the other three
-8-
290
conventional curve features (Figure 4). There is an overlap in the distribution of Fk2 with the
distributions of Fk3 and Fk4 which indicates that Fk3 and Fk4 would not improve the ability to
distinguish the tumor type if Fk2 was known. For the pharmacokinetic parameters, kep and
ve could effectively differentiate the malignant from the benign tumors. The
distribution of Ktrans and vp in malignant and benign tumors (Figure 5) were not
295
separable and it could be not useful to improve the diagnosis performance. Our results
suggest that a combination of all parametric features from the pharmacokinetic model is
significantly better than a combination of all conventional features from the FCM clustering
kinetic curve for predicting the benignity and malignancy in CAD system.
300
4.2 ROC Analysis
The ROC area index AZ over the testing output values was examined to evaluate the
overall performance of the proposed method [21]. The AZ values of the representative
characteristic TIC extracted by FCM clustering were analyzed by combination of the four
conventional curve parameters and four pharmacokinetic parameters (Figure 6). Comparison
305
of the AZ values between different groups of features was made with the z-test. There was
better performance from using Tofts pharmacokinetic model parameters (0.9154) than by
using conventional curve parameters (0.8673) for the FCM clustering extracted TIC though
the difference did not reach statistical significance (p= 0.2331).
310
4.3 Diagnostic Accuracy
The diagnostic performance of 3-D DCE-MRI classification using the binary logistic
regression is illustrated in Table 2 The results show that the accuracy (86.36%), sensitivity
(85.51%) and NPV (84.62%) of combination of the four Tofts pharmacokinetic parameters
were better than the combination of conventional kinetic curve parameters (accuracy 84.85%,
315
sensitivity 79.71%, and NPV 80.28%) while the specificity (87.30%) and PPV (88.06%) were
slightly worse than conventional analysis (specificity 90.48%, PPV 90.16%), (Table 2). The
results from the TIC selected by the interpreting radiologists are shown in Table 3. According
to the classification of the TIC [3], Type I is progressive increased enhancement, type II
-9-
plateau enhancement, and type III early peak and rapid wash out. The distribution of kinetic
320
curve types which belong to benign and malignant tumors is shown in Table 3, and the
performance of these curves using the initial judgment of the interpreting radiologists is
shown in Table 4. If type II is considered a malignant feature, the results from the interpreting
radiologists (accuracy 84.09%, sensitivity 100%, specificity 66.67%, PPV 76.67%, NPV
100%) were better insensitivity and NPV but much worse in accuracy, specificity and PPV
325
compared to the results from using conventional curve characteristics and pharmacokinetic
parameters. The discrepancy in the results might be related to the area of highest clinical
concern for the interpreting radiologists being different from the area extracted by the FCM
clustering technique. Some benign lesions showed a wash out pattern (type III) and plateau
pattern (type II), while some malignant lesions showed progressive increased enhancement
330
(type I), as shown in Table 3.
5. DISCUSSION
Due to the heterogeneity of tumor composition and tremendous amount of 3-D volume data,
manual selection of a ROI in DCE-MRI may fail to pick up the most representative TIC of the
335
tumor based on visual inspection. Therefore, an automatic or semiautomatic method for
DCE-MRI data analysis may be of value to assist clinical interpretation of breast MRI [22].
Color mapping with voxel-based TIC can demonstrate the heterogeneous composition of
breast mass lesions. The final malignancy result of the breast mass with many voxel-based
TIC will be dominated by the TIC component with most relative enhancement. Though
340
representative curve analysis may lose the information of heterogeneous composition, using
FCM clustering technique was shown to increase the malignant kinetic appearance of
cancerous lesion while overlooking that of benign lesions [7]. Many investigations have
shown the relationship between pharmacokinetic modeling of DCE MRI and angiogenesis in
breast cancer [21,23]. In our study, it demonstrated the additional information and value of
345
using pharmacokinetic model to increase diagnostic accuracy of CAD in DCE breast MRI.
In our pilot study, tumor segmentation using the combined kinetic color mapping and
AUC color map has the advantages of highlighting the most significant area representing
- 10 -
tumor with greatest enhancement and localizing VOI for imaging processing in a shorter time
compared to processing the whole 3D DCE-MRI data set. Our approach emphasized tumor
350
segmentation of the significant enhancement component which is associated with physiologic
meaning and correlated with the pharmacokinetic analysis [13]. This approach might be
useful to segment functioning or enhancing pixel in DCE- MRI.
According to prior studies, some quantitative matrices are commonly extracted from
kinetic DCE-MRI, such as initial enhancement rate, maximal enhancement rate and amplitude,
355
and enhancement rate at various time points [7]. These quantitative features are usually not
included as interpretive criteria because of their complexity, ambiguity in visual perception,
and variation of different MR techniques. In a recent study, a significant decrease in false
positive readings of breast MRI has been shown when using a threshold enhancement from a
commercially available CAD system [6]. Using a threshold method to highlight the voxels
360
with significantly increased enhancement can enable total evaluation of the kinetic curves of
all voxels of the whole breast and, therefore, has the benefit to demonstrate tumor
heterogeneity. However, the judgment of lesion category relies on that individual experience
and significant inter-observer variation exists. FCM is a method of clustering which allows
one piece of data to belong to two or more clusters. FCM has been used for automatic
365
segmentation and generation of a clustered concentration data set in DCE-MRI [24]. Using
FCM clustering to extract a representative TIC has been shown to be better than simple
averaging over the entire lesion in DCE-MRI of the breast [7]. In our study, a single
representative TIC of a segmented mass was automatically obtained after placing a seed in a
VOI, instead of many signal time curves from all pixels in a manually selected mass, as in a
370
prior study [6]. Our approach could simplify the result from a representative TIC after fitting
with the Tofts pharmacokinetic model, with an acceptable sensitivity and accuracy (both
approximately 86%), better than conventional curve analysis and thus more suitable for the
screening purpose due to relatively high sensitivity. Voxel-based TICs of the tumor are
possibly quite variable due to their heterogeneous composition, while the FCM clustering
375
technique can produce one single representative TIC. In the future development of CAD for
DCE-MRI, we believe that using a combination of the information from TICs of all voxels
- 11 -
and the FCM clustering extracted TIC of the segmented mass might complement each other to
obtain an even better result for lesion characterization.
In clinical practice, several representative ROIs are usually selected by radiologists in the
380
most enhancing or suspicious region [3]. In contrast, FCM clustering is used to cluster all
curves into several groups and the center of each group indicates the most significant curve
vector to represent this tumor. The discrepancy between the representative TIC from FCM
clustering and that selected by interpreting radiologists was difficult to compare due to
significant overlap in the curve patterns among benign and malignant mass-like lesions as
385
well as influence of lesion morphology during manual selection of TIC by radiologists as
shown in our study. However, there is no doubt that an additional automatic system is of help
to assist radiologists in making a better choice and a more precise diagnosis.
Though different patient groups were enrolled, the results obtained in this study using
the pharmacokinetic model to fit the FCM clustering derived curve without inclusion of
390
morphologic features showed equivalent or even better performance compared with combined
morphologic, conventional TIC features (AZ = 0.9154 vs. AZ = 0.86) in a prior study [8]. It is
therefore reasonable to anticipate that with the combination of this new technique with
morphological features, the diagnostic performance using this proposed system can be further
increased. As compared with the representative curve selected by interpreting radiologists in
395
our study, the sensitivity and NPV of the expert’s selection were higher than our CAD system
(100% and 100% vs. 85.51% and 84.62%) but the accuracy, specificity and PPV were lower
(84.09%, 66.67% and 76.67% vs. 86.36%, 87.30% and 88.06%) (Table 2, Table 4). It was
presumed that FCM clustering might not identify the same representative TIC as selected by
expert radiologists. In view of knowing information of clinical history, lesion morphology and
400
other breast imaging findings, radiologists might adjust their ROI to find out a more
suspicious enhancing region of breast masses to avoid underestimation of any possible
malignancy. Therefore, a combination of the information of morphologic features should be
an important issue to improve the current pilot CAD system with the use of TIC information.
With combination of both conventional kinetic and morphologic information from DCE
405
breast MRI, computer aided diagnosis was reported further improved to distinguish between
- 12 -
non-invasive and invasive lesions and discriminating between metastatic and nonmetastatic
lesions [25]. It is expected to improve the performance of CAD system if the features of
pharmacokinetic model in FCM clustering TIC are added.
Our results of using Tofts model parameters showed distinct differences between
410
benign and malignant lesions in Table 1. There were significantly higher values of
Ktrans and ve in benign lesions while higher values of kep and vp were found in
malignant lesions. According to Furman-Haran E et al [17], highest values of Ktran, kep
and ve were found in invasive ductal carcinomas, followed by DCIS and benign
lesions while fibroadenoma group demonstrated highest value of ve using histogram
415
analysis of parametric mapping. There were more heterogeneous components of either
benign or malignant group in our study. The possible explanations included the
difference of subgroup categorization and methods of quantifications.
There were some limitations of this pilot study. First, this system was semiautomatic
with the requirement of operator to select the most important VOI from integrated color
420
mapping of kinetic curve and AUC. This approach could fasten the speed of computation but
might lose some important information such as mildly enhanced breast tumor like ductal
carcinoma in situ (DCIS). A wider range of thresholding should be applied to include more
mildly enhanced in the future. Second, no morphologic information of segmented breast
mass lesions will be a drawback of the results. However, the performance of our design was at
425
least equivalent to or slightly better than prior report using both morphologic information and
conventional curve analysis [8]. It is reasonably presumed that combined pharmacokinetic
information of the representative TIC using FCM clustering technique will be a better
approach.
In conclusion, our study found that combined pharmacokinetic parametric analysis of
430
the representative TIC extracted from FCM clustering can provide a more accurate approach
to DCE-MRI with a reasonably acceptable result, when compared with conventional kinetic
analytic method. With more and more popular incorporation of CAD diagnosis [7,26] in the
clinical practice, this new approach can potentially be implemented and combined with
- 13 -
morphologic features to increase the diagnostic performance of CAD system for DCE-MRI of
435
the breast.
- 14 -
REFERENCES
[1]
440
[2]
445
[3]
[4]
450
[5]
[6]
455
[7]
[8]
460
[9]
465
[10]
[11]
Kuhl CK, Schrading S, Bieling HB, Wardelmann E, Leutner CC, Koenig R, et
al. MRI for diagnosis of pure ductal carcinoma in situ: a prospective
observational study. Lancet 2007;370:485-492.
Warner E, Plewes DB, Shumak RS, Catzavelos GC, Di Prospero LS, Yaffe MJ,
et al. Comparison of breast magnetic resonance imaging, mammography, and
ultrasound for surveillance of women at high risk for hereditary breast cancer.
J Clin Oncol 2001;19:3524-3531.
Kuhl CK, Mielcareck P, Klaschik S, Leutner C, Wardelmann E, Gieseke J, et
al. Dynamic breast MR imaging: are signal intensity time course data useful
for differential diagnosis of enhancing lesions? Radiology 1999;211:101-110.
Liu PF, Debatin JF, Caduff RF, Kacl G, Garzoli E, Krestin GP. Improved
diagnostic accuracy in dynamic contrast enhanced MRI of the breast by
combined quantitative and qualitative analysis. Br J Radiol 1998;71:501-509.
Nunes LW, Schnall MD, Orel SG. Update of breast MR imaging architectural
interpretation model. Radiology 2001;219:484-494.
Williams TC, DeMartini WB, Partridge SC, Peacock S, Lehman CD. Breast
MR imaging: Computer-aided evaluation program for discriminating benign
from malignant lesions. Radiology 2007;244:94-103.
Chen W, Giger ML, Bick U, Newstead GM. Automatic identification and
classification of characteristic kinetic curves of breast lesions on DCE-MRI.
Med Phys 2006;33:2878-2887.
Chen W, Giger ML, Lan L, Bick U. Computerized interpretation of breast
MRI: investigation of enhancement-variance dynamics. Med Phys
2004;31:1076-1082.
Hayton P, Brady M, Tarassenko L, Moore N. Analysis of dynamic MR breast
images using a model of contrast enhancement. Med Image Anal
1997;1:207-224.
Orel SG, Schnall MD, LiVolsi VA, Troupin RH. Suspicious breast lesions: MR
imaging with radiologic-pathologic correlation. Radiology 1994;190:485-493.
Tofts PS. Modeling tracer kinetics in dynamic Gd-DTPA MR imaging. Jmri-J
Magn Reson Im 1997;7:91-101.
[12]
470
475
Tofts PS, Brix G, Buckley DL, Evelhoch JL, Henderson E, Knopp M, et al.
Estimating kinetic parameters from dynamic contrast-enhanced T-1-weighted
MRI of a diffusable tracer: Standardized quantities and symbols. J Magn
Reson Imaging 1999;10:223-232.
[13] Walker-Samuel S, Leach MO, Collins DJ. Evaluation of response to treatment
using DCE-MRI: the relationship between initial area under the gadolinium
curve (IAUGC) and quantitative pharmacokinetic analysis. Phys Med Biol
2006;51:3593-3602.
- 15 -
480
[14]
Wiener JI, Schilling KJ, Adami C, Obuchowski NA. Assessment of suspected
breast cancer by MRI: A prospective clinical trial using a combined kinetic
and morphologic analysis. AJR Am J Roentgen 2005;184:878-886.
[15]
Galbraith SM, Lodge MA, Taylor NJ, Rustin GJS, Bentzen S, Stirling JJ, et al.
Reproducibility of dynamic contrast-enhanced MRI in human muscle and
tumours: comparison of quantitative and semi-quantitative analysis. NMR in
Biomedicine 2002;15:132-142.
Hojjatoleslami SA, Kittler J. Region growing: A new approach. IEEE Trans
Image Process 1998;7:1079-1084.
Pham DL, Prince JL. An adaptive fuzzy C-means algorithm for image
segmentation in the presence of intensity inhomogeneities. Pattern Recogn
Lett 1999;20:57-68.
Marquardt DW. An algorithm for least-squares estimation of nonlinear
[16]
485
[17]
[18]
490
[19]
[20]
495
[21]
500
[22]
[23]
505
[24]
[25]
510
515
parameters. J Soc Ind Appl Math 1963;11:431-441.
Field AP. Discovering statistics using SPSS. Los Angeles: SAGE Publications:
2009. 821 p.
Xiao Y, Hua J, Dougherty ER. Quantification of the impact of feature selection
on the variance of cross-validation error estimation. EURASIP J Bioinform
Syst Biol 2007:16354.
Tofts PS, Brix G, Buckley DL, Evelhoch JL, Henderson E, Knopp MV, et al.
Estimating kinetic parameters from dynamic contrast-enhanced T1-weighted
MRI of a diffusable tracer: standardized quantities and symbols. J Magn
Reson Imaging 1999;10:223-232.
Turnbull LW. Dynamic contrast-enhanced MRI in the diagnosis and
management of breast cancer. NMR Biomed 2009;22:28-39.
Chang YC, Huang CS, Liu YJ, Chen JH, Lu YS, Tseng WY. Angiogenic
response of locally advanced breast cancer to neoadjuvant chemotherapy
evaluated with parametric histogram from dynamic contrast-enhanced MRI.
Phys Med Biol 2004;49:3593-3602.
Li J, Yu Y, Zhang Y, Bao S, Wu C, Wang X, et al. A clinically feasible method
to estimate pharmacokinetic parameters in breast cancer. Med Phys
2009;36:3786-3794.
Bhooshan N, Giger ML, Jansen SA, Li H, Lan L, Newstead GM. Cancerous
breast lesions on dynamic contrast-enhanced MR images: computerized
characterization for image-based prognostic markers. Radiology
2010;254:680-690.
[26] Wang LC, DeMartini WB, Partridge SC, Peacock S, Lehman CD.
MRI-detected suspicious breast lesions: predictive values of kinetic features
measured by computer-aided evaluation. AJR Am J Roentgenol
2009;193:826-831.
- 16 -
FIGURE LEGENDS
520
Figure 1.
525
Figure 2.
530
Figure 3.
535
540
545
Illustration of the 3-D segmentation procedure for an invasive ductal
carcinoma at a representative slice. The kinetic color map represents the
distribution of the maximum enhancement in the DCE study. The AUC
color map represents the largest cumulated signal intensity over time in
the time-signal intensity curve (TIC). The integrated color map, which is
a combined color map of the yellow part of the kinetic color map and the
red part of the AUC color map. After obtaining user selected volume of
interest (VOI), tumor of interest is then applied for extracting tumor in
DCE MRI.
Two examples of invasive ductal carcinoma. The tumors are segmented
step by step. (a) The kinetic color map. (b) The AUC color map. (c) The
integrated color map with VOI (the green region), and (d) the contour of
segmented tumor (the red line) on the enhanced source image at the
fourth phase of the dynamic contrast enhanced study.
Feature extraction of a representative time-signal intensity curve (TIC) in
an invasive ductal carcinoma using the fuzzy C mean (FCM) clustering
technique. (a) A heterogeneous segmented breast cancer marked with
three points for TIC analysis. (b) Three different types of TIC (A, B, and
C) due to heterogeneous tumor compositions. (c) The representative
kinetic curve identified by the FCM clustering shows a characteristic
rapid slope and wash-out pattern suggestive of malignancy.
Figure 4. The scatter plots of the conventional curve characteristics. (a) Fk2 vs. Fk1,
(b) Fk2 vs. Fk3, and (c) Fk2 vs. Fk4, based on the fuzzy c-means (FCM)
clustering..
Figure 5. The scatter plots of the Tofts curve characteristics. (a) Ktrans vs. kep, (b) kep
vs. vp, and (c) vp vs. ve, based on the fuzzy c-means (FCM) clustering
Figure 6. Comparison between the receiver operating curves (ROC) from using
conventional curve characteristics and Tofts pharmacokinetic model
parameters for the representative time-signal intensity curve (TIC) extracted
by the fuzzy c-means (FCM) clustering technique.
- 17 -
550
Table 1 Comparison among different parameters, derived from conventional classification
and pharmacokinetic models, of kinetic curves obtained from DCE-MRI for
discriminating benign from malignant masses based on the fuzzy c-means (FCM)
clustering.
Method
Parameter
s
characteristics
-
Malignant
1.589±0.452
-
Benign
-
465.372
Malignant
-
180.144
Benign
-
0.003
Malignant
-
0.009
Benign
-
4.914×1
0-4
(min-1)
Malignant
-
8.969×1
0-4
Ktrans
Benign
-
0.860
(min-1)
Malignant
-
0.652
Benign
-
0.142
Malignant
-
0.411
Benign
0.415±0.162
-
Malignant
0.600±0.215
-
(min)
Fk3
(min-1)
Fk4
kep
(min-1)
Tofts model
parameters
vp
Benign
ve
555
*
Median
1.233±0.681
Fk2
curve
SD
Benign
Fk1
Conventional
Mean ±
Type
p-value
<0.001
<0.001
<0.001
0.005
<0.001
<0.001
<0.001
5.639
Malignant
<0.001
1.438
The conventional curve characteristics are maximum enhancement (Fk1), time to peak
(Fk2), uptake rate (Fk3), and washout rate (Fk4) while the Tofts pharmacokinetic model
parameters are Ktrans, kep, vp, and veb representing re-parameterizations of the Tofts
variables.
Ktrans, kep, vp, and ve are respectively volume transfer constant between
plasma and EES, rate constant between plasma and EES, fractional volume of plasma
560
and fractional volume of EES.
**
The mean value, standard deviation (SD), median value, and p-value of Student’s t test
or Mann-Whitney U test for various features in the benign and malignant cases.
- 18 -
Table 2.
565
Comparison of the performance of using combined conventional curve characters
and combined pharmacokinetic parameters of the representative time-signal
intensity curve (TIC) extracted by using fuzzy c-means (FCM) clustering in the
diagnosis of mass-like breast lesion.
Item
Combined
conventional curve
pharmacokinetic
characteristics*
parameters*
86.36
(114/132)
Accuracy
84.85
(112/132)
Sensitivity
79.71
(55/69)
Specificity
PPV
NPV
570
Combined
90.48
(57/63)
90.16
(55/61)
80.28
(57/71)
85.51
(59/69)
87.30
(55/63)
88.06
(59/67)
84.62
(55/65)
Note: The test results of combined conventional curve characteristics and combined
pharmacokinetic parameters of the representative time-signal intensity curve (TIC)
extracted with fuzzy c-means (FCM) clustering were tested by binary logistic regression
with the leave-one-out cross-validation method. TP, true positive; TN, true negative; FP,
false positive; FN, false negative.
575
Accuracy = (TP+TN) / (TP+TN+FP+FN)
Sensitivity = TP / (TP+FN)
Specificity = TN / (TN+FP)
Positive Predictive Value = TP / (TP+FP)
Negative Predictive Value = TN / (TN+FN)
580
- 19 -
Table 3. The distribution of curve types in benign and malignant breast masses according to
interpreting radiologists.
Type I
Type II
Type III
Benign (N=46)
42
18
3
Malignant (N=78)
0
8
61
585
- 20 -
Table 4. The performance of the time-signal intensity curve (TIC) selected by interpreting
radiologists.
590
Item
TIC Categorization by Radiologists
Accuracy
84.09% (111/132)
Sensitivity
100% (69/69)
Specificity
66.67% (42/63)
PPV
76.67% (69/90)
NPV
100% (42/42)
Note that both type II (plateau enhancement) and type III (early peak enhancement
followed by wash out) are considered malignant.
- 21 -